The Burry debate put GPU depreciation on every board agenda. Real secondary-market data on how H100, A100, and V100 values actually decayed — and what that means for refresh timing. · Published July 26, 2026 · Industrial Surplus World desk
Ever since the Burry short thesis made GPU depreciation a board-level argument, CFOs have been asking the same question: what do these accelerators actually lose per year? The accounting schedules say 3–5 years straight-line. The secondary market says something more interesting — and more useful for refresh planning.
Real transaction data shows GPU values decay in steps, not slopes. An H100 SXM that traded near $30,000 in early 2024 holds $18,000–$30,000 as of July 26, 2026 — far better than straight-line assumptions — because inference demand created a durable floor under Hopper. But each supply event (Blackwell volume shipments, cloud price cuts, a hyperscaler fleet release) steps the whole curve down at once. A100 80GB tells the same story one generation later: still $9,000–$19,000, five years after launch, but repricing downward at every B200 milestone.
The practical implication for finance teams: residual value is a timing decision, not a constant. Fleets sold ahead of a generation ramp consistently capture 20–40% more than fleets sold twelve months into one. That gap usually exceeds the marginal compute value extracted by holding — which is the entire sell-now-or-hold calculation in one sentence.
Our public price index tracks these curves weekly with condition tiers per SKU, and the B200 vs H100 analysis covers the current step in detail. For a live number against your actual serials, call (954) 488-0700 — written offers land the same business day.